ScienceClaw
🔬🦞 A self-evolving AI research colleague for scientists. 285 skills, zero hallucination, persistent memory.
Install / Use
/learn @beita6969/ScienceClawREADME
Why ScienceClaw?
General-purpose AI assistants are built for everyone. ScienceClaw is built for researchers.
The core idea is simple: an AI that does real scientific work — searching literature, querying databases, running analyses — and gets better at it the more you use it. It remembers your research context across sessions, adapts its skills to your field, and never fabricates a citation.
ScienceClaw is built on the OpenClaw engine, but redesigned from the ground up for academic research.
<p align="center"> <img src="assets/comparison.png" alt="ScienceClaw vs Standard AI" width="720" /> </p>🧬 Core 1: Self-Evolving Skills
This is ScienceClaw's most important feature.
Most AI tools ship with a fixed set of capabilities. ScienceClaw's skills evolve with you. Every time you complete a research task, the system learns:
<p align="center"> <img src="assets/skill-evolution.png" alt="Skill Self-Evolution Cycle" width="720" /> </p>What this means in practice:
- Week 1: You study immunology. ScienceClaw learns that PubMed + Semantic Scholar works best for your queries, that you prefer forest plots over tables, and that you always need PMID + DOI in citations.
- Week 4: The system has created specialized skills for your subfield — optimized search templates, preferred statistical methods, database priority chains tuned to immunology literature.
- Month 3: ScienceClaw handles your domain like a trained research assistant. It knows which databases to hit first, which journals matter, and how you like your output formatted.
Compared to standard OpenClaw: OpenClaw ships with ~54 general-purpose skills that don't change. ScienceClaw starts with 285 skills and grows from there — the agent writes new
SKILL.mdfiles at runtime without any redeployment.
🧠 Core 2: Research Memory That Persists
Standard AI assistants forget everything when the conversation ends. ScienceClaw doesn't.
<p align="center"> <img src="assets/memory-layers.png" alt="Four-Layer Research Memory" width="720" /> </p>What this enables:
- "Continue the literature review we started last Tuesday" — it remembers where you left off
- "Use the same search strategy that worked for the BRCA2 project" — it retrieves past patterns
- Cross-session knowledge accumulation — findings from project A can inform project B
- Smart context pruning — when the context window fills up, it preserves statistical results, effect sizes, and key citations while compacting intermediate steps
Compared to standard OpenClaw: OpenClaw has a basic memory plugin. ScienceClaw adds temporal decay weighting, LanceDB vector storage, and cross-session research pattern retrieval — specifically designed for long-running academic work.
⏱️ Core 3: Built for Long-Duration Research
A real literature review takes hours, not seconds. Most AI tools time out after a few minutes. ScienceClaw is engineered for extended research sessions:
| Capability | Standard OpenClaw | ScienceClaw | | ------------------- | ---------------------- | ----------------------------------------------------------------- | | Agent timeout | 600s (10 min) | 3600s (1 hour+) | | Session persistence | Ends with conversation | Heartbeat keeps sessions alive across interruptions | | Research depth | Single-pass response | Multi-phase protocol with mandatory depth thresholds | | Minimum effort | No guarantee | Quick=5, Survey=30, Review=60, Systematic=100+ tool calls | | Early stopping | Common | Anti-premature-conclusion checklist blocks shallow answers | | Context management | Basic truncation | Smart compaction preserves key findings when context fills up |
The persistence protocol enforces real research depth. Before ScienceClaw concludes any task, it must verify:
- ✅ Searched at least 3 different databases/sources
- ✅ Retrieved full metadata (not just titles)
- ✅ Cross-referenced findings across sources
- ✅ Checked for contradictory evidence
- ✅ Verified key statistics against primary sources
- ✅ Organized results into a structured output file
- ✅ Met the minimum tool-call threshold for the task type
If any box is unchecked, it keeps working instead of giving you a half-baked answer.
Compared to standard OpenClaw: OpenClaw's default 10-minute timeout is fine for sending messages and setting reminders. ScienceClaw's 1-hour sessions with heartbeat monitoring and mandatory depth enforcement are built for real academic research.
🚫 Core 4: Zero Hallucination
This is the highest-priority rule in the entire system. It's non-negotiable.
The problem: General AI assistants routinely fabricate citations — inventing DOIs, making up author names, citing papers that don't exist. In scientific work, this is catastrophic.
ScienceClaw's approach:
EVERY citation must come from a tool result in the CURRENT conversation.
If a database didn't return it → you can't cite it.
If you're not sure → say "not verified" explicitly.
If you can't find evidence → say so. Don't guess.
No "I think." No "probably." No hallucinated PMIDs.
This is enforced at the protocol level in SCIENCE.md — the 629-line research protocol that governs all agent behavior. It's not a suggestion. It's a hard rule that applies before any other instruction.
Compared to standard OpenClaw: OpenClaw has no special hallucination controls. ScienceClaw's SCIENCE.md protocol treats every factual claim as requiring evidence — the same standard you'd apply to a manuscript under peer review.
🌍 Core 5: All of Science, Not Just Biomedicine
ScienceClaw covers natural sciences AND social sciences across dozens of disciplines:
<p align="center"> <img src="assets/disciplines.png" alt="Scientific Discipline Coverage" width="720" /> </p> <details> <summary><strong>📋 Full discipline & database list</strong></summary>Natural Sciences
| Domain | Key Skills & Databases | | ------------------------- | --------------------------------------------------------------------- | | Biomedicine | PubMed, UniProt, KEGG, PDB, ClinicalTrials, gnomAD, scanpy, biopython | | Chemistry | PubChem, ChEMBL, RDKit, drug-discovery, molecular-dynamics | | Genomics | NCBI Entrez, Ensembl, ClinVar, GEO, phylogenetics | | Materials Science | Materials Project, pymatgen, materials-screening | | Physics | astropy, quantum-computing, physics-solver, simulation | | Environmental Science | Copernicus climate data, geospatial analysis, GIS tools | | Food Science | Specialized analysis pipelines |
Social Sciences
| Domain | Key Skills & Databases | | --------------------- | ------------------------------------------------------- | | Economics | World Bank, SSRN, census data, econometrics | | Political Science | Policy analysis, legislative data | | Psychology | Experimental design, statistical testing, meta-analysis | | Linguistics | spaCy, NLTK, NLP analysis | | Education | Research methodology, assessment analysis | | Sociology | Network analysis, survey methods |
Cross-Disciplinary Tools
| Category | Capabilities | | ----------------- | ----------------------------------------------------------------------------------------------------- | | Statistics | SciPy, statsmodels, scikit-learn, effect sizes, confidence intervals, multiple comparison corrections | | Visualization | matplotlib, plotly, seaborn, publication-quality figures | | Writing | LaTeX papers, systematic reviews (PRISMA), grant proposals, patent drafting | | Mathematics | SymPy symbolic computation, numerical methods, optimization |
</details>285 skills total — and growing, because the self-evolution system creates new ones as you work.
Compared to standard OpenClaw: OpenClaw has no scientific database integrations. No PubMed, no UniProt, no arXiv, no World Bank. ScienceClaw connects to 25+ academic databases with structured API query skills across all major scientific disciplines.
Quick Start
# Clone
git clone https://github.com/beita6969/ScienceClaw.git
cd ScienceClaw
# One-click setup (installs everything: Node, Python, MCP servers, skills)
chmod +x setup.sh && ./s
Related Skills
Hook Development
80.2kThis skill should be used when the user asks to "create a hook", "add a PreToolUse/PostToolUse/Stop hook", "validate tool use", "implement prompt-based hooks", "use ${CLAUDE_PLUGIN_ROOT}", "set up event-driven automation", "block dangerous commands", or mentions hook events (PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart, SessionEnd, UserPromptSubmit, PreCompact, Notification). Provides comprehensive guidance for creating and implementing Claude Code plugin hooks with focus on advanced prompt-based hooks API.
MCP Integration
80.2kThis skill should be used when the user asks to "add MCP server", "integrate MCP", "configure MCP in plugin", "use .mcp.json", "set up Model Context Protocol", "connect external service", mentions "${CLAUDE_PLUGIN_ROOT} with MCP", or discusses MCP server types (SSE, stdio, HTTP, WebSocket). Provides comprehensive guidance for integrating Model Context Protocol servers into Claude Code plugins for external tool and service integration.
Plugin Structure
80.2kThis skill should be used when the user asks to "create a plugin", "scaffold a plugin", "understand plugin structure", "organize plugin components", "set up plugin.json", "use ${CLAUDE_PLUGIN_ROOT}", "add commands/agents/skills/hooks", "configure auto-discovery", or needs guidance on plugin directory layout, manifest configuration, component organization, file naming conventions, or Claude Code plugin architecture best practices.
Skill Development
80.2kThis skill should be used when the user wants to "create a skill", "add a skill to plugin", "write a new skill", "improve skill description", "organize skill content", or needs guidance on skill structure, progressive disclosure, or skill development best practices for Claude Code plugins.
